Explaining and Predicting the Effects of Neurostimulation via Neuronal 1 Excitation/Inhibition

26 Previous research has highlighted the role of the excitation/inhibition ratio (E/I) for typical 27 and atypical development, mental health, cognition, and learning. Parallel research has 28 highlighted the benefits of high-frequency random noise stimulation (tRNS)—an excitatory 29 form of neurostimulation—on learning. We examined the E/I as a potential mechanism and 30 whether tRNS effect on learning depends on E/I as measured by the aperiodic exponent. In 31 addition to manipulating E/I using tRNS, we also manipulated the level of learning 32 (learning/overlearning) that has been shown to influence E/I. One hundred and two 33 participants received either sham stimulation or 20 min DLPFC tRNS during a mathematical 34 learning task. We showed a lower aperiodic exponent, which reflects higher E/I, after tRNS, 35 and that higher baseline aperiodic exponent, which reflects lower E/I, predicted greater 36 benefit from tRNS specifically for the learning task. In contrast to previous MRS-based E/I 37 studies, we found no effect of the learning manipulation on E/I. Our results highlight the role 38 of E/I as a marker for neurostimulation efficacy and learning. This mechanistic understanding 39 could provide stronger potential to augment learning. At the same time, we offer new insights 40 on the quantification of E/I using EEG vs MRS to foster better theoretical understanding and 41 its utilisation in future research. 42 45 46 47 48 49 our findings indicate that this involvement may not be due to E/I alterations relating to skill acquisition, but due to the participants’ E/I level at baseline.

oscillatory activity. The power of aperiodic activity decreases exponentially with increasing 83 frequency (see Figure 1A), and relates to the negative slope in log-log space (see Figure 1B). 84 In contrast to the previous assumption that aperiodic activity reflects a background noise in 85 the EEG spectrum, accumulating evidence shows the importance of aperiodic activity in 86 understanding brain functions and behavior. Also, periodic activity has been shown to be 87 confounded due to misestimating spectral power since participants vary in center frequencies 88 if a predefined spectral range is applied (Lansbergen et al., 2011). Therefore, Donoghue et al. 89 (2020) recommend to parameterize neural power spectra by also analyzing the aperiodic 90 activity in the spectrum. Aperiodic activity consists of an aperiodic exponent that can be 91 defined as x in a 1/f x function, which reflects the previously mentioned negative slope in log-92 log space and, thus, the pattern of power across frequencies. The exponent of aperiodic 93 activity is thought to underlie the integration of underlying synaptic currents (Buzsáki et al.,94 2013), and a likely mechanism of changes in the aperiodic exponent has been linked to the E/I 95 of field potentials shown by EEG recordings (Gao et al., 2017). A higher E/I relates to a lower 96 aperiodic exponent and vice versa. The power of inhibitory GABA currents leads to a rapid 97 decay in the power spectrum at higher frequencies, and, thus a steeper (negatively sloped) 98 exponent (see Figure 1C). The opposite happens for excitatory currents, where power is 99 stable for lower frequencies and declines more slowly for higher frequencies, which is shown 100 Aperiodic Activity 5 in a flatter (closer to zero) exponent (see Figure 1C). Shortly, the higher the E/I, the lower the 101 exponent value (see Figure 1D).  . 115 Recent MRS findings have linked better mathematical skills to higher E/I in young 116 adults and the reverse in younger participants (Zacharopoulos et al., 2021b). Moreover, MRS-117 based E/I can predict future mathematical reasoning (Zacharopoulos et al., 2021c). Since 118 previous studies report a link between E/I and mathematical achievement, we applied tRNS 119 while participants solved arithmetic multiplications during a mathematical learning paradigm. 120 Furthermore, we manipulated the degree of skill acquisition to induce learning or 121 overlearning. Based on previous studies, we defined learning as practising a skill during 122 performance improvement (i.e., before learning plateaus) and overlearning as the point after 123 performance improvement when a plateau has been reached (Shibata et al., 2017). Learning 124 and overlearning have been linked to an increase and a decrease in E/I, respectively, in an 125 MRS study (Shibata et al., 2017). 126 We aimed to impact E/I directly using tRNS, as well as indirectly by manipulating the 127 level of skill acquisition (learning/overlearning) to examine whether: 1) tRNS will increase 128 E/I as measured by aperiodic exponent; 2) The direction of change in aperiodic exponent 129 between pre-and posttest depends on the learning condition: it decreases in the learning 130 condition and increases in the overlearning condition; 3) tRNS efficacy on a 131 learning/overlearning task depends on the individual baseline aperiodic exponent. That is, the 132 tRNS-induced reduction of the aperiodic exponent differs across participants, depending on 133 their baseline aperiodic exponent (i.e., E/I levels) (Krause et al., 2013).    The Impact of tRNS and Learning on the Aperiodic Exponent

174
To further explain the significant effect of stimulation, we plotted the aperiodic 175 exponent change for each stimulation group separately (see Figure 3 and Figure S2 for  However, it should be noted that in contrast to our expectations, type of task 181 (learning/overlearning) did not influence the aperiodic exponent change from pre to post. Because a frequentist approach does not allow to accept the null result (i.e., no effect 191 of task on aperiodic exponent), we reran the same ANCOVA using a Bayesian approach on 192 the aperiodic exponent change. Our results, as presented in the Supplementary Information, 193 strengthens the conclusion that tRNS impacts the aperiodic exponent, while task has no effect.

194
These findings match the idea that tRNS leads to higher excitation and therefore a lower (i.e., 195 flatter) aperiodic exponent and that such effect is independent of learning/overlearning.

216
We checked the posterior distributions that captures the uncertainty surrounding the 217 magnitude of an effect. Typically, a posterior distribution higher or equal to 75% (below or 218 above zero) is chosen as a threshold to indicate that an effect is present. The choice for a 219 certain cutoff criterion depends on the potential risks and benefits of the intervention (i.e., 220 Ahn et al., 2018), and in this context it means that there is a 75% chance that the alternative 221 hypothesis (i.e., the presence of an effect) is true. Figure 4A shows that there is a 90% 222 probability that tRNS lowers median RTs on average during both tasks and thus improves 223 performance (see Figure S3 for all main effects). The most important effect is the three-way 224 interaction between tRNS, task, and aperiodic exponent at baseline, which was the model with 225 th best fit (see Figure 4 and Figure S3). Notably, the posterior probability of the presence of 226 a three-way interaction between tRNS X task X baseline aperiodic exponent is 82% (see 227 Figure 4B).
To understand the source of this 3-way interaction we dissected it by running the 229 model for learning and overlearning separately (see Figure 4C). For the learning task the 230 posterior distribution for the interaction between stimulation and the baseline aperiodic 231 exponent was 96%. We therefore further dissected the model for sham and tRNS separately 232 for the main effect of the baseline aperiodic exponent in the learning task. We found that 233 those with higher baseline aperiodic exponent performed worse than those with a lower 234 exponent in the sham condition (posterior distribution=77%). However, this effect was 235 reversed when tRNS was applied, showing better performance for those with a higher 236 baseline aperiodic exponent (posterior distribution=88%).

237
Oppositely for the overlearning task, the posterior distribution was 56% indicating no 238 support for an interaction between stimulation and baseline aperiodic exponent in this task.   and baseline aperiodic exponent that shows that there is a 82% probability that this interaction  sham stimulation group (for statistical details see Table S1). Also, no difference was found 273 between the groups in the impact of these sensations on their subjective performance. The aim of the present study was to impact E/I (measured by means of the aperiodic 276 exponent) directly using tRNS, and indirectly by manipulating the level of skill acquisition 277 (learning/overlearning). The aperiodic exponent decreased after tRNS, indicating an increased 278 E/I, which corroborates with the working mechanisms of tRNS as a stimulation method that 279 increases neuronal excitation. However, we found no effect of task manipulation on the 280 aperiodic exponent, which indicates that overlearning a skill does not affect the aperiodic 281 exponent or the E/I respectively. This was in contrast to our expectation formed by a previous 282 MRS study (Shibata et al., 2017). We also showed that tRNS efficacy on learning depends on 283 the baseline aperiodic exponent, but no beneficical effects were found for overlearning. To

299
Our findings suggest a working mechanism of tRNS efficacy, related to E/I. Whether this mechanism is dependent or independent from stochastic resonance is a question for further 301 research. 302 We did not find an effect of task manipulation (i.e., mathematical 303 learning/overlearning) on the aperiodic exponent. It is likely that the aperiodic exponent was adults (see Supplementary Information). We found that higher glutamate/GABA measured 365 with MRS (i.e., higher E/I) was significantly associated with an increased aperiodic exponent 366 (i.e., lower E/I) over the left IPS. No relation was found for the left MFG (see Figure S4).

367
This shows that MRS and EEG may measure different aspects of E/I. These preliminary  post resting state (rs)-EEG recording (more than 25% of their data were rejected) (see Table 1 394 for demographic data). The final sample (n=75) was composed of 22 participants in the sham- Open Science Framework (see https://osf.io/y4xar). However, the analyses (i.e., neuronal 402 avalanches) presented in this preregistration did not yield significant results (see Figure S6). 403 We later came across the work on the aperiodic exponent as a measure of E/I, which we used 404 in this study. We investigated whether the participants in the four conditions (i.e., sham-learning,  Table S2). Every multiplication 428 problem presented consisted of two-digit times one-digit operands with a two-digit answer (e.g.,  Participants were instructed to wear a headphone to cancel out any surrounding noise, and 441 to speak clearly and loudly in the microphone without mumbling or clearing the throat (e.g.,

442
saying "eh-em", which would be registered as a response). Lastly, participants were informed 443 that there was no time limit for answering, and they were urged to avoid errors. In total, 180 multiplication problems were administered in both the overlearning and the 476 learning condition (see Table S3). The structure of the tasks was identical to the baseline task 477 (see Figure 5A). Both conditions consisted of 18 blocks, comprised of the same number of 478 trials and were presented in a fixed order. After three blocks, participants had a one-minute 479 break. Therefore, in the learning condition, a subset of ten problems was presented once in each 480 block. In the overlearning condition, a subset of five problems was selected, which was presented twice in each block (i.e., less information to learn in comparison to the learning 482 condition). This manipulation allowed us to use the same task and duration yet influencing the defined by the international 10-20 system for EEG recording (see Figure 5C). Two Pistim

489
Ag/AgCl electrodes were used with a 1 cm radius and a surface area of 3.14 cm 2 each. A current 78%; active tRNS: 79%; χ(1)=0.03, p=1). This independent data is supported by our data, which 506 did not find differences in sensations between both groups (see Results section).

507
Electrophysiological Data 508 Rs-EEG recordings were made before baseline allocation and at the end of the experiment 509 as stated in our preregistration (see Figure 5B). Electrophysiological data were obtained with 510 32 Ag/AgCl electrodes according to the international 10/20 EEG system using the wireless and any data recorded before the presentation of the fixation point was removed. Every data 524 file was manually checked and high-amplitude artefacts due to muscle movement, sweating or 525 electrode malfunction were rejected. After pre-processing, Independent Component Analysis 526 (ICA) was performed to remove stereotyped artefacts such as eye movements (e.g., blinks), 527 heart rate activity and muscular activity. A maximum of six components per data file were 528 removed, and a maximum of five bad channels were interpolated. EEG segments that contained 529 artefacts that could not be removed by ICA were visually inspected and rejected from the analysis (Delorme et al., 2007). If more than 25 percent of the rs-EEG data were rejected after 531 pre-processing and ICA, data of the subject were discarded from analysis.

533
The rs-EEG data of the remaining participants were separated in 2-second segments assess alertness, good-bad mood, tiredness, calmness, and restlessness.

545
Then, participants completed an rs-EEG pre-measurement of eight minutes where they 546 were informed to sit as still as possible (see Figure 5B). Then, a training was presented that 547 consisted of four different arithmetic multiplications. Hereafter, the baseline task was started 548 and a variance minimization procedure (based on response times) followed the baseline task in 549 a double-blind fashion to determine which participants would be allocated to which group (see 550 https://osf.io/y4xar for a detailed explanation of this procedure). This procedure is superior to 551 random assignment for assigning participants to groups before an intervention (Sella et al., task. More information about the procedure of these tasks, which is beyond the scope of the present manuscript, can be found on https://osf.io/y4xar. At the end of the behavioral tasks, an 556 8 minute rs-EEG measurement was recorded.

557
Behavioral Data Cleaning 558 We excluded responses below 200 ms due to possible noises picked up by the 559 microphone or mumbling of the subject (0.89% for the baseline task, 2.74% for the learning 560 task, and 7.07% for the overlearning task). We also excluded wrong responses from the baseline 561 ability task (16.32%), the learning task (10.65%), and the overlearning task (7.7%). We  We originally used glmer for our analysis but due to model complexity we revert to brms. All Bayesian models were ran with 5000 iterations and 4 chains each, and used dummy coding. We 581 made the decision to not look at the RTs on trial level in our Bayesian models due to the high 582 amount of introduced noise (i.e., variance between trials) and the lack of computational power.

583
Due the right skewness of the median response time, the lognormal family was used. Also, all 584 continuous independent variables were centered to prevent multicollinearity. 585 586